Few-Shot Graph Anomaly Detection via Dual-Level Knowledge Distillation.

Area of Science:

• Graph analytics
• Machine learning
• Data science

Background:

• Graph anomaly detection is vital across many fields, but acquiring labeled data is challenging and costly.
• Few-shot learning offers a solution by requiring minimal labeled data, yet conventional models often underutilize auxiliary information.
• Suboptimal performance in existing few-shot graph anomaly detection methods stems from incomplete exploitation of auxiliary datasets.

Purpose of the Study:

• To propose DualKD, a dual-level knowledge distillation approach for graph anomaly detection.
• To enhance the generalization capabilities of few-shot learning models in graph anomaly detection.
• To address the limitations of conventional few-shot models in leveraging auxiliary data.

Main Methods:

• A teacher model is trained to generate soft labels (prediction distributions) capturing data uncertainty.
• A student model is trained using these soft labels to learn detailed node features.
• Dual representation distillation (short and long) is employed to transfer knowledge from auxiliary to target sets.

Main Results:

• DualKD significantly outperforms advanced baseline methods in graph anomaly detection.
• The proposed method demonstrates enhanced identification performance across four experimental datasets.
• Knowledge distillation effectively improves the generalization capabilities of few-shot graph anomaly detection.

Conclusions:

• DualKD offers a superior approach to few-shot graph anomaly detection by effectively utilizing dual-level knowledge distillation.
• The method successfully transfers knowledge, leading to improved performance compared to existing techniques.
• DualKD provides a robust solution for scenarios with limited labeled data in graph anomaly detection.